Add JAX PPO ladder verification pass

configs/jax_ppo/gate3-exploiter.yaml

@@ -0,0 +1,40 @@
+run:
+  experiment_name: jax-ppo-gate3-exploiter
+  seed: 20260705
+  learner_seat: 0
+  total_updates: 250
+  log_every: 1
+  checkpoint_every: 10
+  artifact_root: /mnt/2tbhdd/coolrl-lost-cities-artifacts/verification
+
+opponent:
+  name: gate3_checkpoint_frozen
+
+network:
+  hidden_size: 512
+  num_layers: 3
+
+ppo:
+  batch_games: 8192
+  rollout_steps: 400
+  gamma: 1.0
+  gae_lambda: 0.95
+  clip_epsilon: 0.2
+  entropy_coef: 0.01
+  value_coef: 0.5
+  max_grad_norm: 0.5
+  learning_rate: 0.0003
+  epochs: 4
+  minibatches: 128
+
+reward:
+  terminal_scale: 50.0
+  potential_shaping_initial: 1.0
+  potential_shaping_final: 0.0
+  potential_shaping_anneal_steps: 5000000
+
+evaluation:
+  games: 10000
+  duplicate: true
+  shuffle_bank_seed: 20260704
+  batch_games: 8192

docs/reports/verification-pass-2026-07-04.md

@@ -0,0 +1,191 @@
+# Verification Pass - 2026-07-04
+
+**Status:** FAIL - stop before league self-play.
+**Implementation start commit:** `bb52ef9`.
+**Primary checkpoint under audit:** `/mnt/2tbhdd/coolrl-lost-cities-artifacts/jax-ppo-static-opponents/2026-07-04_230150_jax-ppo-cautious/latest`.
+**Verification artifacts:** `/mnt/2tbhdd/coolrl-lost-cities-artifacts/verification/2026-07-04_phase1/`.
+
+## Decision
+
+Do not start snapshot-pool league self-play from the current ladder artifacts.
+The duplicate/Wilson protocol is reproducible, and the obvious seat-perspective
+bug was not found, but the cautious anchor is not a trustworthy permanent Elo
+anchor: it stalls heavily in mirror play and gives the trained policy a very
+clean win while the trained policy opens almost all colors.
+
+## Transcript Audit
+
+Transcript file:
+
+```text
+docs/reports/verification-pass-2026-07-04-transcripts.txt
+```
+
+Rendered protocol: 10 fixed shuffle-bank deck orders, duplicate seat-swapped,
+for 20 total games of gate-3 checkpoint vs `heuristic_cautious`.
+
+Summary from the transcript dump:
+
+| Metric | Value |
+| --- | ---: |
+| Games | 20 |
+| Agent mean score diff | +144.500 |
+| Agent score diff min/max | -93 / +230 |
+| Agent opened colors mean | 5.000 |
+| Mean game length | 159.200 |
+| Length min / p50 / p95 / max | 63 / 109 / 400 / 400 |
+| Max-steps rate | 0.150 |
+| Handshake play events | 86 |
+| Cautious openings | 78 |
+| Cautious low openings below rank 7 | 3 |
+| Cautious discards immediately playable by opponent | 162 |
+
+Interpretation: the sample confirms the concern behind this pass. The agent is
+not winning by a restrained 2-3 color expert pattern; it opens all 5 colors on
+average and still wins by a wide margin. The cautious policy is also leaking
+many immediately useful discard tops.
+
+## Protocol Recheck
+
+The original ladder evaluation was audited in code. The `eval` path uses:
+
+- explicit shuffle banks generated from `evaluation.shuffle_bank_seed`;
+- duplicate seat-swapped loops over learner seat 0 and learner seat 1;
+- Wilson confidence intervals from aggregate wins.
+
+The protocol itself is not the main failure. The failure is that the original
+gate summary omitted `max_steps_rate` and score-diff distribution canaries.
+
+Recomputed gate-3 duplicate evaluation with distribution canaries:
+
+```text
+/mnt/2tbhdd/coolrl-lost-cities-artifacts/verification/2026-07-04_phase1/gate3_reverification_distribution.json
+```
+
+| Metric | Value |
+| --- | ---: |
+| Games | 20,000 |
+| Win rate | 0.95960 |
+| Wilson 95% | [0.95678, 0.96224] |
+| Mean score diff | +142.9178 |
+| Score diff p05 / p50 / p95 | +21 / +152 / +227 |
+| Opened colors/game | 4.98965 |
+| Positive expeditions/game | 4.09470 |
+| Play action rate | 0.27639 |
+| Mean game length | 185.3179 |
+| Game length p50 / p95 / max | 115 / 400 / 400 |
+| Max-steps rate | 0.24070 |
+
+The 24.07% forced-end rate is a high-severity canary failure for using this
+gate as a clean league baseline.
+
+## Perspective And Mirror Tests
+
+New tests added:
+
+- hand-authored heuristic behavior checks for weak unopened hands, strong
+  balanced opens, and P1 own-board perspective;
+- static-policy duplicate mirror checks for all three heuristic policies;
+- gate-3 checkpoint duplicate self-mirror check, skipped on CPU-only runs and
+  executed under CUDA JAX locally.
+
+Commands run:
+
+```bash
+uv run pytest -q tests/lost_cities_jax/test_ppo_stack.py
+uv run --with 'jax[cuda12]' pytest -q \
+  tests/lost_cities_jax/test_ppo_stack.py::test_gate3_checkpoint_duplicate_self_mirror_score_diff_is_zero
+```
+
+Results:
+
+- CPU focused tests: 11 passed, 1 skipped.
+- CUDA checkpoint self-mirror test: 1 passed.
+
+Mirror evaluation artifacts:
+
+```text
+/mnt/2tbhdd/coolrl-lost-cities-artifacts/verification/2026-07-04_phase1/mirror_discard_only.json
+/mnt/2tbhdd/coolrl-lost-cities-artifacts/verification/2026-07-04_phase1/mirror_heuristic_balanced.json
+/mnt/2tbhdd/coolrl-lost-cities-artifacts/verification/2026-07-04_phase1/mirror_heuristic_cautious.json
+/mnt/2tbhdd/coolrl-lost-cities-artifacts/verification/2026-07-04_phase1/mirror_gate3_checkpoint.json
+```
+
+| Mirror policy | Mean diff | Win/loss symmetry | Max-steps rate | Play action rate | Mean length |
+| --- | ---: | --- | ---: | ---: | ---: |
+| `discard_only` | 0.0 | exact | 0.0000 | 0.0000 | 44.0000 |
+| `heuristic_balanced` | 0.0 | exact | 0.4683 | 0.0422 | 231.1210 |
+| `heuristic_cautious` | 0.0 | exact | 0.8956 | 0.0053 | 368.8435 |
+| gate-3 checkpoint | 0.0 | exact | 0.0000 | 0.7839 | 45.4292 |
+
+Interpretation: the broad P0/P1 perspective bug is unlikely. Duplicate mirror
+score differences cancel exactly. The serious issue is heuristic quality:
+`heuristic_cautious` mirror play almost never opens expeditions and reaches
+forced termination in 89.56% of games.
+
+## Exploiter Baseline
+
+Frozen opponent: gate-3 checkpoint. New exploiter: random init PPO, same 8192 x
+400 rollout shape, 250 updates, shaping annealed by the standard schedule.
+
+Training run:
+
+```text
+/mnt/2tbhdd/coolrl-lost-cities-artifacts/verification/2026-07-04_235948_jax-ppo-gate3-exploiter/
+```
+
+Final train-row canaries:
+
+| Metric | Value |
+| --- | ---: |
+| return_mean | +0.24118 |
+| play_action_rate | 0.71362 |
+| opened_colors_mean | 4.91016 |
+| positive_expeditions_mean | 2.07922 |
+| game_length_mean | 49.12524 |
+| max_steps_rate | 0.00000 |
+
+Duplicate exploiter evaluation:
+
+```text
+/mnt/2tbhdd/coolrl-lost-cities-artifacts/verification/2026-07-04_phase1/exploiter_vs_gate3_duplicate.json
+```
+
+| Metric | Value |
+| --- | ---: |
+| Games | 20,000 |
+| Exploiter win rate | 0.67115 |
+| Wilson 95% | [0.66461, 0.67763] |
+| Mean score diff | +24.1371 |
+| Score diff p05 / p50 / p95 | -50 / +21 / +110 |
+| Opened colors/game | 4.89710 |
+| Positive expeditions/game | 2.12865 |
+| Play action rate | 0.74612 |
+| Mean game length | 48.0017 |
+| Max-steps rate | 0.00000 |
+
+This is now the measured exploitability baseline. It is not by itself a stop
+condition, but it reinforces that the current gate-3 checkpoint is not a
+finished robust policy.
+
+## Findings
+
+| Severity | Finding | Evidence | Consequence |
+| --- | --- | --- | --- |
+| High | `heuristic_cautious` is a stalling anchor, not a strong cautious baseline. | Mirror max-steps rate 89.56%, play action rate 0.0053. | Do not use it as a permanent Elo anchor without fixing/replacing it. |
+| High | Gate-3 result is contaminated by forced-end games and over-opening. | Gate-3 max-steps rate 24.07%, opened colors/game 4.98965. | The ladder PASS remains reproducible, but it should not become a league baseline. |
+| Medium | Cautious discards many immediately playable cards to the opponent. | Transcript sample: 162 such discards in 20 games. | The agent may be exploiting discard leakage rather than learning robust play. |
+| Medium | A same-budget PPO exploiter beats the gate-3 checkpoint. | Exploiter duplicate win rate 67.115%, mean diff +24.1371. | League success should reduce this number, but Phase 2 should wait for anchor repair. |
+| Low | No broad seat perspective bug found. | Static and checkpoint duplicate mirrors have exact zero mean diff. | Seat symmetry is not the likely explanation for the ladder result. |
+
+## Stop Condition
+
+Phase 1 did not pass cleanly. Per the work order, Phase 2 is not started.
+
+Recommended next work:
+
+1. Replace or repair `heuristic_balanced` and `heuristic_cautious` so mirror
+   play has near-zero forced-end rate and realistic play/open rates.
+2. Add `max_steps_rate` and game-length quantiles to every gate report.
+3. Rerun the static ladder gates with the repaired anchors before initializing
+   a snapshot-pool league.

scripts/render_jax_ppo_transcripts.py

@@ -0,0 +1,334 @@
+"""Render human-readable JAX PPO Lost Cities transcripts."""
+
+from __future__ import annotations
+
+import argparse
+import json
+from collections import Counter
+from pathlib import Path
+from typing import NamedTuple
+
+import jax
+import jax.numpy as jnp
+import numpy as np
+
+from lost_cities_jax.engine import (
+    board_score,
+    current_hand_sorted,
+    decode_action,
+    reset_from_order,
+    step,
+)
+from lost_cities_jax.opponents import policy_by_name
+from lost_cities_jax.ppo import checkpoint_policy, load_config, make_shuffle_bank
+from lost_cities_jax.types import (
+    CARDS_PER_COLOR,
+    DISCARD,
+    DRAW_DECK,
+    LOC_P0_BOARD,
+    MAX_STEPS,
+    N_CARDS,
+    N_COLORS,
+    PLAY,
+    State,
+)
+
+COLOR_NAMES = ["red", "green", "white", "blue", "yellow"]
+STEP_JIT = jax.jit(step)
+
+
+class PolicySpec(NamedTuple):
+    label: str
+    fn: object
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--config", required=True)
+    parser.add_argument("--checkpoint", required=True)
+    parser.add_argument("--opponent", default="heuristic_cautious")
+    parser.add_argument("--shuffle-bank-seed", type=int, default=20260704)
+    parser.add_argument("--pairs", type=int, default=10)
+    parser.add_argument("--output", required=True)
+    args = parser.parse_args()
+
+    cfg = load_config(args.config)
+    agent = PolicySpec("gate3_checkpoint", jax.jit(checkpoint_policy(cfg, args.checkpoint)))
+    opponent = PolicySpec(args.opponent, jax.jit(policy_by_name(args.opponent)))
+    orders = make_shuffle_bank(args.shuffle_bank_seed, args.pairs)
+
+    sections = []
+    summaries = []
+    for idx, order in enumerate(orders):
+        sections.append(f"\n\n## Pair {idx:02d} / learner seat 0\n")
+        text, summary = render_game(order, [agent, opponent], agent_seat=0, game_id=idx * 2)
+        sections.append(text)
+        summaries.append(summary)
+
+        sections.append(f"\n\n## Pair {idx:02d} / learner seat 1\n")
+        text, summary = render_game(order, [opponent, agent], agent_seat=1, game_id=idx * 2 + 1)
+        sections.append(text)
+        summaries.append(summary)
+
+    output = render_summary(summaries, args) + "".join(sections)
+    Path(args.output).parent.mkdir(parents=True, exist_ok=True)
+    Path(args.output).write_text(output, encoding="utf-8")
+
+
+def render_game(
+    deck_order: list[int], policies: list[PolicySpec], *, agent_seat: int, game_id: int
+) -> tuple[str, dict]:
+    state = reset_from_order(jnp.asarray(deck_order, dtype=jnp.int8))
+    lines = [f"Game {game_id}: P0={policies[0].label}, P1={policies[1].label}"]
+    handshake_events = []
+    cautious_openings = []
+    cautious_useful_discards = []
+
+    while not bool(state.done):
+        player = int(state.to_move)
+        policy = policies[player]
+        key = jax.random.PRNGKey(game_id * 1000 + int(state.step_count))
+        hand = [int(x) for x in current_hand_sorted(state, player)]
+        action = int(policy.fn(state, jnp.asarray(player, dtype=jnp.int32), key))
+        hand_slot, place_type, draw_source = [int(x) for x in decode_action(jnp.asarray(action))]
+        card = hand[hand_slot]
+        drawn = peek_drawn_card(state, place_type, draw_source)
+        before_col_len = int(state.col_len[player, card_color(card)])
+        before_hand = [card_label(x) for x in hand if x < N_CARDS]
+        before_scores = [int(x) for x in board_score(state)]
+
+        if policy.label == "heuristic_cautious":
+            if place_type == PLAY and before_col_len == 0:
+                cautious_openings.append(
+                    {
+                        "ply": int(state.step_count),
+                        "seat": player,
+                        "card": card_label(card),
+                        "rank": card_rank(card),
+                    }
+                )
+            if place_type == DISCARD and can_play_card(state, 1 - player, card):
+                cautious_useful_discards.append(
+                    {
+                        "ply": int(state.step_count),
+                        "seat": player,
+                        "card": card_label(card),
+                        "opponent": 1 - player,
+                    }
+                )
+
+        if place_type == PLAY and is_handshake(card):
+            handshake_events.append(
+                {
+                    "ply": int(state.step_count),
+                    "seat": player,
+                    "policy": policy.label,
+                    "card": card_label(card),
+                    "hand": before_hand,
+                }
+            )
+
+        next_state, _, _ = STEP_JIT(state, jnp.asarray(action, dtype=jnp.int32))
+        place_text = "PLAY" if place_type == PLAY else "DISCARD"
+        target = "expedition" if place_type == PLAY else "discard"
+        draw_text = "deck" if draw_source == DRAW_DECK else f"{COLOR_NAMES[draw_source - 1]} pile"
+        lines.append(
+            f"Ply {int(state.step_count):03d} P{player} {policy.label}: "
+            f"{place_text} {card_label(card)} to {target}; "
+            f"DRAW {draw_text} -> {card_label(drawn)}; "
+            f"score_before={before_scores}"
+        )
+        lines.extend(render_board(next_state))
+        state = next_state
+
+    final_scores = [int(x) for x in board_score(state)]
+    lines.append(f"Final score: P0={final_scores[0]} P1={final_scores[1]}")
+    lines.extend(render_final_breakdown(state))
+    agent_score = final_scores[agent_seat]
+    opponent_score = final_scores[1 - agent_seat]
+    summary = {
+        "game_id": game_id,
+        "agent_seat": agent_seat,
+        "agent_score_diff": agent_score - opponent_score,
+        "agent_opened_colors": int(jnp.sum(state.col_len[agent_seat] > 0)),
+        "length": int(state.step_count),
+        "max_steps": int(state.step_count) >= MAX_STEPS,
+        "handshake_events": handshake_events,
+        "cautious_openings": cautious_openings,
+        "cautious_useful_discards": cautious_useful_discards,
+    }
+    return "\n".join(lines) + "\n", summary
+
+
+def render_summary(summaries: list[dict], args: argparse.Namespace) -> str:
+    diffs = np.asarray([item["agent_score_diff"] for item in summaries], dtype=np.float64)
+    opened = np.asarray([item["agent_opened_colors"] for item in summaries], dtype=np.float64)
+    lengths = np.asarray([item["length"] for item in summaries], dtype=np.float64)
+    handshakes = [event for item in summaries for event in item["handshake_events"]]
+    cautious_openings = [event for item in summaries for event in item["cautious_openings"]]
+    useful_discards = [event for item in summaries for event in item["cautious_useful_discards"]]
+    opening_ranks = Counter(str(event["rank"]) for event in cautious_openings)
+    low_openings = [
+        event for event in cautious_openings if event["rank"] not in ("HS",) and event["rank"] < 7
+    ]
+
+    lines = [
+        "# Gate-3 Transcript Dump",
+        "",
+        f"config: `{args.config}`",
+        f"checkpoint: `{args.checkpoint}`",
+        f"opponent: `{args.opponent}`",
+        f"duplicate pairs: `{args.pairs}`",
+        "",
+        "## Summary",
+        "",
+        f"games: {len(summaries)}",
+        f"agent_score_diff_mean: {float(np.mean(diffs)):.3f}",
+        f"agent_score_diff_min_max: {int(np.min(diffs))} / {int(np.max(diffs))}",
+        f"agent_opened_colors_mean: {float(np.mean(opened)):.3f}",
+        f"game_length_mean: {float(np.mean(lengths)):.3f}",
+        f"game_length_min_p50_p95_max: {int(np.min(lengths))} / "
+        f"{float(np.quantile(lengths, 0.50)):.1f} / "
+        f"{float(np.quantile(lengths, 0.95)):.1f} / {int(np.max(lengths))}",
+        f"max_steps_rate: {float(np.mean([item['max_steps'] for item in summaries])):.3f}",
+        f"handshake_play_events: {len(handshakes)}",
+        f"cautious_openings: {len(cautious_openings)}",
+        f"cautious_opening_rank_counts: {json.dumps(dict(sorted(opening_ranks.items())))}",
+        f"cautious_low_openings_lt7: {len(low_openings)}",
+        f"cautious_discards_immediately_playable_by_opponent: {len(useful_discards)}",
+        "",
+        "## Handshake Play Contexts",
+        "",
+    ]
+    if handshakes:
+        for event in handshakes[:80]:
+            lines.append(
+                f"- game_event ply={event['ply']} seat=P{event['seat']} "
+                f"policy={event['policy']} card={event['card']} hand={event['hand']}"
+            )
+        if len(handshakes) > 80:
+            lines.append(f"- ... {len(handshakes) - 80} more")
+    else:
+        lines.append("- none")
+
+    lines.extend(["", "## Cautious Opening Audit", ""])
+    if cautious_openings:
+        for event in cautious_openings[:80]:
+            lines.append(
+                f"- ply={event['ply']} seat=P{event['seat']} card={event['card']} "
+                f"rank={event['rank']}"
+            )
+        if len(cautious_openings) > 80:
+            lines.append(f"- ... {len(cautious_openings) - 80} more")
+    else:
+        lines.append("- none")
+
+    lines.extend(["", "## Cautious Useful Discard Audit", ""])
+    if useful_discards:
+        for event in useful_discards[:80]:
+            lines.append(
+                f"- ply={event['ply']} seat=P{event['seat']} card={event['card']} "
+                f"opponent=P{event['opponent']}"
+            )
+        if len(useful_discards) > 80:
+            lines.append(f"- ... {len(useful_discards) - 80} more")
+    else:
+        lines.append("- none")
+    lines.append("")
+    return "\n".join(lines)
+
+
+def render_board(state: State) -> list[str]:
+    scores = [int(x) for x in board_score(state)]
+    lines = [f"  Board scores: P0={scores[0]} P1={scores[1]}"]
+    for player in range(2):
+        parts = []
+        for color in range(N_COLORS):
+            cards = board_cards(state, player, color)
+            score = color_score(cards)
+            parts.append(f"{COLOR_NAMES[color]}={format_cards(cards)}({score:+d})")
+        lines.append(f"  P{player}: " + " | ".join(parts))
+    return lines
+
+
+def render_final_breakdown(state: State) -> list[str]:
+    lines = ["Final color breakdown:"]
+    for player in range(2):
+        parts = []
+        for color in range(N_COLORS):
+            cards = board_cards(state, player, color)
+            parts.append(f"{COLOR_NAMES[color]} {format_cards(cards)} => {color_score(cards):+d}")
+        lines.append(f"  P{player}: " + "; ".join(parts))
+    return lines
+
+
+def board_cards(state: State, player: int, color: int) -> list[int]:
+    loc = np.asarray(state.card_loc)
+    board_loc = LOC_P0_BOARD + player
+    start = color * CARDS_PER_COLOR
+    cards = [card for card in range(start, start + CARDS_PER_COLOR) if int(loc[card]) == board_loc]
+    return sorted(
+        cards, key=lambda card: (0 if is_handshake(card) else 1, card_rank_value(card), card)
+    )
+
+
+def color_score(cards: list[int]) -> int:
+    if not cards:
+        return 0
+    handshakes = sum(1 for card in cards if is_handshake(card))
+    ranks = [card_rank_value(card) for card in cards if not is_handshake(card)]
+    score = (sum(ranks) - 20) * (1 + handshakes)
+    if len(cards) >= 8:
+        score += 20
+    return score
+
+
+def peek_drawn_card(state: State, place_type: int, draw_source: int) -> int:
+    del place_type
+    if draw_source == DRAW_DECK:
+        return int(state.deck_order[int(state.draw_ptr)])
+    color = draw_source - 1
+    length = int(state.pile_len[color])
+    return int(state.pile[color, length - 1])
+
+
+def can_play_card(state: State, player: int, card: int) -> bool:
+    top = int(state.col_top[player, card_color(card)])
+    if is_handshake(card):
+        return top == 0
+    return card_rank_value(card) > top
+
+
+def format_cards(cards: list[int]) -> str:
+    if not cards:
+        return "[]"
+    return "[" + ",".join(card_label(card) for card in cards) + "]"
+
+
+def card_label(card: int) -> str:
+    if card < 0 or card >= N_CARDS:
+        return "none"
+    color = COLOR_NAMES[card_color(card)][0].upper()
+    if is_handshake(card):
+        return f"{color}HS{card % CARDS_PER_COLOR + 1}"
+    return f"{color}{card_rank_value(card)}"
+
+
+def card_color(card: int) -> int:
+    return card // CARDS_PER_COLOR
+
+
+def is_handshake(card: int) -> bool:
+    return card % CARDS_PER_COLOR < 3
+
+
+def card_rank(card: int) -> int | str:
+    return "HS" if is_handshake(card) else card_rank_value(card)
+
+
+def card_rank_value(card: int) -> int:
+    return card % CARDS_PER_COLOR - 1
+
+
+if __name__ == "__main__":
+    main()

src/lost_cities_jax/ppo.py

@@ -147,6 +147,16 @@ class EvalBatch(NamedTuple):
     games: jax.Array
 
 
+class MatchBatch(NamedTuple):
+    score_diff: jax.Array
+    game_length: jax.Array
+    play_action_sum: jax.Array
+    learner_action_sum: jax.Array
+    opened_colors: jax.Array
+    positive_expeditions: jax.Array
+    max_steps: jax.Array
+
+
 def load_config(path: str | Path | None = None, overrides: dict | None = None) -> JaxPPOConfig:
     data: dict = {}
     if path is not None:
@@ -179,7 +189,12 @@ def create_train_state(cfg: JaxPPOConfig, rng: jax.Array) -> TrainState:
     return TrainState.create(apply_fn=model.apply, params=params, tx=tx)
 
 
-def train(cfg: JaxPPOConfig, *, resume: str | None = None) -> Path:
+def train(
+    cfg: JaxPPOConfig,
+    *,
+    resume: str | None = None,
+    opponent_policy=None,
+) -> Path:
     run_dir = _create_run_dir(cfg)
     _write_json(run_dir / "config.json", asdict(cfg))
     metrics_path = run_dir / "metrics.jsonl"
@@ -192,7 +207,8 @@ def train(cfg: JaxPPOConfig, *, resume: str | None = None) -> Path:
     if resume:
         state = restore_checkpoint(Path(resume), state)
 
-    opponent_policy = policy_by_name(cfg.opponent.name)
+    if opponent_policy is None:
+        opponent_policy = policy_by_name(cfg.opponent.name)
     train_iteration = make_train_iteration(cfg, opponent_policy)
     start = time.perf_counter()
 
@@ -584,6 +600,246 @@ def make_eval_batch_fn(cfg: JaxPPOConfig, opponent_policy):
     return eval_batch
 
 
+def checkpoint_policy(cfg: JaxPPOConfig, checkpoint: str | Path):
+    state = create_train_state(cfg, jax.random.PRNGKey(0))
+    state = restore_checkpoint(Path(checkpoint), state)
+    return checkpoint_policy_from_params(cfg, state.params)
+
+
+def checkpoint_policy_from_params(cfg: JaxPPOConfig, params):
+    model = ActorCritic(cfg.network.hidden_size, cfg.network.num_layers)
+
+    def policy(state: State, player: jax.Array, rng: jax.Array) -> jax.Array:
+        del rng
+        obs = observation(state, player)[None, :]
+        legal = legal_action_mask(state)[None, :]
+        logits, _ = model.apply(params, obs)
+        return jnp.argmax(mask_logits(logits, legal), axis=-1)[0].astype(jnp.int32)
+
+    return policy
+
+
+def train_against_checkpoint(
+    cfg: JaxPPOConfig,
+    opponent_cfg: JaxPPOConfig,
+    opponent_checkpoint: str | Path,
+    *,
+    resume: str | None = None,
+) -> Path:
+    policy = checkpoint_policy(opponent_cfg, opponent_checkpoint)
+    return train(cfg, resume=resume, opponent_policy=policy)
+
+
+def evaluate_checkpoint_match(
+    learner_cfg: JaxPPOConfig,
+    learner_checkpoint: str | Path,
+    opponent_cfg: JaxPPOConfig,
+    opponent_checkpoint: str | Path,
+    *,
+    games: int | None = None,
+    duplicate: bool | None = None,
+    output: str | Path | None = None,
+) -> dict:
+    learner_policy = checkpoint_policy(learner_cfg, learner_checkpoint)
+    opponent_policy = checkpoint_policy(opponent_cfg, opponent_checkpoint)
+    games = games or learner_cfg.evaluation.games
+    duplicate = learner_cfg.evaluation.duplicate if duplicate is None else duplicate
+    return evaluate_policy_match(
+        learner_policy,
+        opponent_policy,
+        games=games,
+        duplicate=duplicate,
+        shuffle_bank_seed=learner_cfg.evaluation.shuffle_bank_seed,
+        batch_games=learner_cfg.evaluation.batch_games,
+        output=output,
+    )
+
+
+def evaluate_checkpoint_vs_static(
+    learner_cfg: JaxPPOConfig,
+    learner_checkpoint: str | Path,
+    opponent_name: str,
+    *,
+    games: int | None = None,
+    duplicate: bool | None = None,
+    output: str | Path | None = None,
+) -> dict:
+    learner_policy = checkpoint_policy(learner_cfg, learner_checkpoint)
+    opponent_policy = policy_by_name(opponent_name)
+    games = games or learner_cfg.evaluation.games
+    duplicate = learner_cfg.evaluation.duplicate if duplicate is None else duplicate
+    result = evaluate_policy_match(
+        learner_policy,
+        opponent_policy,
+        games=games,
+        duplicate=duplicate,
+        shuffle_bank_seed=learner_cfg.evaluation.shuffle_bank_seed,
+        batch_games=learner_cfg.evaluation.batch_games,
+        output=output,
+    )
+    result["opponent"] = opponent_name
+    return result
+
+
+def evaluate_static_mirror(
+    policy_name: str,
+    *,
+    games: int,
+    duplicate: bool = True,
+    shuffle_bank_seed: int = 20260704,
+    batch_games: int = 8192,
+    output: str | Path | None = None,
+) -> dict:
+    policy = policy_by_name(policy_name)
+    return evaluate_policy_match(
+        policy,
+        policy,
+        games=games,
+        duplicate=duplicate,
+        shuffle_bank_seed=shuffle_bank_seed,
+        batch_games=batch_games,
+        output=output,
+    )
+
+
+def evaluate_policy_match(
+    learner_policy,
+    opponent_policy,
+    *,
+    games: int,
+    duplicate: bool,
+    shuffle_bank_seed: int,
+    batch_games: int,
+    output: str | Path | None = None,
+) -> dict:
+    eval_batch = make_policy_match_batch_fn(learner_policy, opponent_policy)
+    diffs = []
+    lengths = []
+    play_actions = 0.0
+    learner_actions = 0.0
+    opened_sum = 0.0
+    positive_sum = 0.0
+    max_steps_sum = 0.0
+
+    for seat in [0, 1] if duplicate else [0]:
+        for start in range(0, games, batch_games):
+            count = min(batch_games, games - start)
+            orders = make_shuffle_bank(shuffle_bank_seed + start, count)
+            batch = eval_batch(jnp.asarray(orders, dtype=jnp.int8), seat)
+            batch = jax.tree_util.tree_map(lambda x: x.block_until_ready(), batch)
+            diffs.append(jax.device_get(batch["score_diff"]))
+            lengths.append(jax.device_get(batch["game_length"]))
+            play_actions += float(jnp.sum(batch["play_action_sum"]))
+            learner_actions += float(jnp.sum(batch["learner_action_sum"]))
+            opened_sum += float(jnp.sum(batch["opened_colors"]))
+            positive_sum += float(jnp.sum(batch["positive_expeditions"]))
+            max_steps_sum += float(jnp.sum(batch["max_steps"]))
+
+    import numpy as np
+
+    diff_arr = np.concatenate(diffs).astype(np.float64)
+    length_arr = np.concatenate(lengths).astype(np.float64)
+    total_games = int(diff_arr.shape[0])
+    wins = int(np.sum(diff_arr > 0))
+    losses = int(np.sum(diff_arr < 0))
+    ties = int(np.sum(diff_arr == 0))
+    wilson_low, wilson_high = wilson_interval(wins, total_games)
+    quantiles = np.quantile(diff_arr, [0.0, 0.05, 0.25, 0.5, 0.75, 0.95, 1.0])
+    hist_counts, hist_edges = np.histogram(
+        diff_arr,
+        bins=[-400, -300, -200, -100, -50, 0, 50, 100, 200, 300, 400, 600],
+    )
+    result = {
+        "duplicate": duplicate,
+        "games": total_games,
+        "wins": float(wins),
+        "losses": float(losses),
+        "ties": float(ties),
+        "win_rate": wins / total_games,
+        "wilson_low": wilson_low,
+        "wilson_high": wilson_high,
+        "mean_score_diff": float(np.mean(diff_arr)),
+        "score_diff_sum": float(np.sum(diff_arr)),
+        "score_diff_quantiles": {
+            "min": float(quantiles[0]),
+            "p05": float(quantiles[1]),
+            "p25": float(quantiles[2]),
+            "p50": float(quantiles[3]),
+            "p75": float(quantiles[4]),
+            "p95": float(quantiles[5]),
+            "max": float(quantiles[6]),
+        },
+        "score_diff_histogram": {
+            "edges": hist_edges.astype(float).tolist(),
+            "counts": hist_counts.astype(int).tolist(),
+        },
+        "mean_game_length": float(np.mean(length_arr)),
+        "game_length_quantiles": {
+            "min": float(np.min(length_arr)),
+            "p50": float(np.quantile(length_arr, 0.50)),
+            "p95": float(np.quantile(length_arr, 0.95)),
+            "max": float(np.max(length_arr)),
+        },
+        "max_steps_rate": max_steps_sum / total_games,
+        "play_action_rate": play_actions / max(learner_actions, 1.0),
+        "opened_colors_per_game": opened_sum / total_games,
+        "positive_expeditions_per_game": positive_sum / total_games,
+    }
+    if output:
+        _write_json(Path(output), result)
+    return result
+
+
+def make_policy_match_batch_fn(learner_policy, opponent_policy):
+    @jax.jit
+    def eval_batch(deck_orders: jax.Array, learner_seat: int):
+        learner = jnp.asarray(learner_seat, dtype=jnp.int32)
+        opponent = 1 - learner
+        env_state = jax.vmap(reset_from_order)(deck_orders)
+        keys = jax.random.split(jax.random.PRNGKey(0), deck_orders.shape[0])
+
+        def body(carry, _):
+            env, key = carry
+            key, learner_key, opponent_key = jax.random.split(key, 3)
+            learner_keys = jax.random.split(learner_key, deck_orders.shape[0])
+            opponent_keys = jax.random.split(opponent_key, deck_orders.shape[0])
+            learner_actions = jax.vmap(learner_policy, in_axes=(0, None, 0))(
+                env, learner, learner_keys
+            )
+            opponent_actions = jax.vmap(opponent_policy, in_axes=(0, None, 0))(
+                env, opponent, opponent_keys
+            )
+            learner_turn = env.to_move.astype(jnp.int32) == learner
+            active = ~env.done
+            actions = jnp.where(learner_turn, learner_actions, opponent_actions)
+            next_env, _, _ = jax.vmap(step, in_axes=(0, 0))(env, actions)
+            place_type = (actions % 12) // 6
+            return (next_env, key), (
+                (place_type == PLAY) & learner_turn & active,
+                learner_turn & active,
+            )
+
+        (final_env, _), (play_actions, learner_actions) = jax.lax.scan(
+            body, (env_state, keys[0]), xs=None, length=MAX_STEPS
+        )
+        scores = jax.vmap(board_score)(final_env)
+        diff = scores[:, learner] - scores[:, opponent]
+        color_scores = jax.vmap(color_scores_for_player, in_axes=(0, None))(final_env, learner)
+        opened = jnp.sum(final_env.col_len[:, learner, :] > 0, axis=-1)
+        positive = jnp.sum(color_scores > 0, axis=-1)
+        return MatchBatch(
+            score_diff=diff,
+            game_length=final_env.step_count,
+            play_action_sum=jnp.sum(play_actions, axis=0),
+            learner_action_sum=jnp.sum(learner_actions, axis=0),
+            opened_colors=opened,
+            positive_expeditions=positive,
+            max_steps=final_env.step_count >= MAX_STEPS,
+        )._asdict()
+
+    return eval_batch
+
+
 def rollout_metrics(
     transitions: Transition, final_env: State, learner: jax.Array
 ) -> dict[str, jax.Array]:
@@ -716,6 +972,13 @@ def cli_main(argv: list[str] | None = None) -> None:
     train_parser.add_argument("--resume")
     train_parser.add_argument("--set", action="append", default=[])
 
+    exploiter_parser = sub.add_parser("train-exploiter")
+    exploiter_parser.add_argument("--config", required=True)
+    exploiter_parser.add_argument("--opponent-config", required=True)
+    exploiter_parser.add_argument("--opponent-checkpoint", required=True)
+    exploiter_parser.add_argument("--resume")
+    exploiter_parser.add_argument("--set", action="append", default=[])
+
     eval_parser = sub.add_parser("eval")
     eval_parser.add_argument("--config", required=True)
     eval_parser.add_argument("--checkpoint", required=True)
@@ -725,17 +988,61 @@ def cli_main(argv: list[str] | None = None) -> None:
     eval_parser.add_argument("--output")
     eval_parser.add_argument("--set", action="append", default=[])
 
+    checkpoint_eval_parser = sub.add_parser("eval-checkpoint")
+    checkpoint_eval_parser.add_argument("--config", required=True)
+    checkpoint_eval_parser.add_argument("--checkpoint", required=True)
+    checkpoint_eval_parser.add_argument("--opponent-config", required=True)
+    checkpoint_eval_parser.add_argument("--opponent-checkpoint", required=True)
+    checkpoint_eval_parser.add_argument("--games", type=int)
+    checkpoint_eval_parser.add_argument("--duplicate", action="store_true")
+    checkpoint_eval_parser.add_argument("--output")
+    checkpoint_eval_parser.add_argument("--set", action="append", default=[])
+
+    checkpoint_static_eval_parser = sub.add_parser("eval-checkpoint-static")
+    checkpoint_static_eval_parser.add_argument("--config", required=True)
+    checkpoint_static_eval_parser.add_argument("--checkpoint", required=True)
+    checkpoint_static_eval_parser.add_argument("--opponent", required=True)
+    checkpoint_static_eval_parser.add_argument("--games", type=int)
+    checkpoint_static_eval_parser.add_argument("--duplicate", action="store_true")
+    checkpoint_static_eval_parser.add_argument("--output")
+    checkpoint_static_eval_parser.add_argument("--set", action="append", default=[])
+
+    static_mirror_parser = sub.add_parser("eval-static-mirror")
+    static_mirror_parser.add_argument("--policy", required=True)
+    static_mirror_parser.add_argument("--games", type=int, default=10_000)
+    static_mirror_parser.add_argument("--batch-games", type=int, default=8192)
+    static_mirror_parser.add_argument("--shuffle-bank-seed", type=int, default=20260704)
+    static_mirror_parser.add_argument("--duplicate", action="store_true")
+    static_mirror_parser.add_argument("--output")
+
     args = parser.parse_args(argv)
-    cfg = load_config(args.config, overrides=parse_overrides(args.set))
+    cfg = (
+        load_config(args.config, overrides=parse_overrides(args.set))
+        if hasattr(args, "config")
+        else None
+    )
     if args.command == "rollout-smoke":
+        assert cfg is not None
         row = random_rollout(cfg)
         print(json.dumps(row, indent=2, sort_keys=True))
         if args.output:
             _write_json(Path(args.output), row)
     elif args.command == "train":
+        assert cfg is not None
         run_dir = train(cfg, resume=args.resume)
         print(run_dir)
+    elif args.command == "train-exploiter":
+        assert cfg is not None
+        opponent_cfg = load_config(args.opponent_config)
+        run_dir = train_against_checkpoint(
+            cfg,
+            opponent_cfg,
+            args.opponent_checkpoint,
+            resume=args.resume,
+        )
+        print(run_dir)
     elif args.command == "eval":
+        assert cfg is not None
         result = evaluate(
             cfg,
             args.checkpoint,
@@ -745,6 +1052,40 @@ def cli_main(argv: list[str] | None = None) -> None:
             output=args.output,
         )
         print(json.dumps(result, indent=2, sort_keys=True))
+    elif args.command == "eval-checkpoint":
+        assert cfg is not None
+        opponent_cfg = load_config(args.opponent_config)
+        result = evaluate_checkpoint_match(
+            cfg,
+            args.checkpoint,
+            opponent_cfg,
+            args.opponent_checkpoint,
+            games=args.games,
+            duplicate=args.duplicate,
+            output=args.output,
+        )
+        print(json.dumps(result, indent=2, sort_keys=True))
+    elif args.command == "eval-checkpoint-static":
+        assert cfg is not None
+        result = evaluate_checkpoint_vs_static(
+            cfg,
+            args.checkpoint,
+            args.opponent,
+            games=args.games,
+            duplicate=args.duplicate,
+            output=args.output,
+        )
+        print(json.dumps(result, indent=2, sort_keys=True))
+    elif args.command == "eval-static-mirror":
+        result = evaluate_static_mirror(
+            args.policy,
+            games=args.games,
+            duplicate=args.duplicate,
+            shuffle_bank_seed=args.shuffle_bank_seed,
+            batch_games=args.batch_games,
+            output=args.output,
+        )
+        print(json.dumps(result, indent=2, sort_keys=True))
 
 
 def _create_run_dir(cfg: JaxPPOConfig) -> Path:
@@ -817,10 +1158,17 @@ __all__ = [
     "ActorCritic",
     "JaxPPOConfig",
     "TrainState",
+    "checkpoint_policy",
+    "checkpoint_policy_from_params",
     "cli_main",
     "create_train_state",
     "evaluate",
+    "evaluate_checkpoint_match",
+    "evaluate_checkpoint_vs_static",
+    "evaluate_policy_match",
+    "evaluate_static_mirror",
     "load_config",
     "random_rollout",
     "train",
+    "train_against_checkpoint",
 ]

tests/lost_cities_jax/test_ppo_stack.py

@@ -1,10 +1,14 @@
 from __future__ import annotations
 
+from pathlib import Path
+
 import jax
 import jax.numpy as jnp
 import numpy as np
+import pytest
 
 from lost_cities_jax import legal_action_mask, reset
+from lost_cities_jax.engine import current_hand_sorted, decode_action, reset_from_order
 from lost_cities_jax.opponents import (
     discard_only_action,
     heuristic_balanced_action,
@@ -18,12 +22,29 @@ from lost_cities_jax.ppo import (
     RunConfig,
     create_train_state,
     evaluate,
+    evaluate_checkpoint_match,
+    evaluate_static_mirror,
     load_config,
     make_train_iteration,
     policy_by_name,
     random_rollout,
     train,
 )
+from lost_cities_jax.types import (
+    CARDS_PER_COLOR,
+    HAND_SIZE,
+    LOC_DECK,
+    LOC_P0_BOARD,
+    LOC_P0_HAND,
+    N_CARDS,
+    N_COLORS,
+    PLAY,
+)
+
+GATE3_CHECKPOINT = (
+    "/mnt/2tbhdd/coolrl-lost-cities-artifacts/jax-ppo-static-opponents/"
+    "2026-07-04_230150_jax-ppo-cautious/latest"
+)
 
 
 def tiny_config(tmp_path) -> JaxPPOConfig:
@@ -49,6 +70,84 @@ def test_static_opponents_return_legal_actions():
         assert mask[action]
 
 
+def test_cautious_discards_weak_unopened_hand():
+    low_cards = [_rank_card(0, rank) for rank in range(2, 7)]
+    state = _manual_state(p0_hand=low_cards, to_move=0)
+    action = int(heuristic_cautious_action(state, jnp.int32(0), jax.random.PRNGKey(1)))
+    _, place_type, _ = [int(x) for x in decode_action(jnp.asarray(action))]
+    assert place_type != PLAY
+
+
+def test_balanced_opens_strong_unopened_rank():
+    strong = _rank_card(0, 10)
+    state = _manual_state(p0_hand=[strong, _rank_card(1, 2), _rank_card(2, 3)], to_move=0)
+    action = int(heuristic_balanced_action(state, jnp.int32(0), jax.random.PRNGKey(1)))
+    hand_slot, place_type, _ = [int(x) for x in decode_action(jnp.asarray(action))]
+    hand = [int(x) for x in current_hand_sorted(state, 0)]
+    assert place_type == PLAY
+    assert hand[hand_slot] == strong
+
+
+def test_cautious_uses_own_board_when_playing_as_p1():
+    existing = _rank_card(0, 5)
+    playable = _rank_card(0, 6)
+    state = _manual_state(
+        p1_hand=[
+            playable,
+            _rank_card(1, 2),
+            _rank_card(1, 3),
+            _rank_card(2, 2),
+            _rank_card(2, 3),
+            _rank_card(3, 2),
+            _rank_card(3, 3),
+            _rank_card(4, 2),
+        ],
+        p1_board=[existing],
+        to_move=1,
+    )
+    action = int(heuristic_cautious_action(state, jnp.int32(1), jax.random.PRNGKey(1)))
+    hand_slot, place_type, _ = [int(x) for x in decode_action(jnp.asarray(action))]
+    hand = [int(x) for x in current_hand_sorted(state, 1)]
+    assert place_type == PLAY
+    assert hand[hand_slot] == playable
+
+
+@pytest.mark.parametrize(
+    "policy_name", ["discard_only", "heuristic_balanced", "heuristic_cautious"]
+)
+def test_static_policy_duplicate_mirror_score_diff_is_zero(policy_name):
+    result = evaluate_static_mirror(
+        policy_name,
+        games=16,
+        duplicate=True,
+        shuffle_bank_seed=20260704,
+        batch_games=16,
+    )
+    assert result["games"] == 32
+    assert result["mean_score_diff"] == 0.0
+    assert result["wins"] == result["losses"]
+
+
+def test_gate3_checkpoint_duplicate_self_mirror_score_diff_is_zero():
+    checkpoint = Path(GATE3_CHECKPOINT)
+    if not checkpoint.exists():
+        pytest.skip(f"gate-3 checkpoint not available: {checkpoint}")
+    if not any(device.platform == "gpu" for device in jax.local_devices()):
+        pytest.skip("gate-3 checkpoint was saved with CUDA sharding; run under CUDA JAX")
+    cfg = load_config("configs/jax_ppo/cautious.yaml")
+    result = evaluate_checkpoint_match(
+        cfg,
+        checkpoint,
+        cfg,
+        checkpoint,
+        games=4,
+        duplicate=True,
+    )
+    assert result["games"] == 8
+    assert result["mean_score_diff"] == 0.0
+    assert result["wins"] == result["losses"]
+
+
 def test_random_rollout_smoke(tmp_path):
     row = random_rollout(tiny_config(tmp_path))
     assert row["env_steps"] == 8 * 16
@@ -84,3 +183,65 @@ def test_load_config_file():
     cfg = load_config("configs/jax_ppo/smoke.yaml")
     assert cfg.opponent.name == "discard_only"
     assert cfg.ppo.batch_games == 64
+
+
+def _rank_card(color: int, rank: int) -> int:
+    return color * CARDS_PER_COLOR + rank + 1
+
+
+def _manual_state(
+    *,
+    p0_hand: list[int] | None = None,
+    p1_hand: list[int] | None = None,
+    p0_board: list[int] | None = None,
+    p1_board: list[int] | None = None,
+    to_move: int = 0,
+):
+    p0_hand = _fill_hand(p0_hand or [], set((p0_board or []) + (p1_board or [])))
+    p1_hand = _fill_hand(p1_hand or [], set(p0_hand + (p0_board or []) + (p1_board or [])))
+    rest = [card for card in range(N_CARDS) if card not in set(p0_hand + p1_hand)]
+    deck_order = jnp.asarray(p0_hand + p1_hand + rest, dtype=jnp.int8)
+    state = reset_from_order(deck_order)
+    card_loc = jnp.full((N_CARDS,), LOC_DECK, dtype=jnp.int8)
+    card_loc = card_loc.at[jnp.asarray(p0_hand)].set(LOC_P0_HAND)
+    card_loc = card_loc.at[jnp.asarray(p1_hand)].set(LOC_P0_HAND + 1)
+    if p0_board:
+        card_loc = card_loc.at[jnp.asarray(p0_board)].set(LOC_P0_BOARD)
+    if p1_board:
+        card_loc = card_loc.at[jnp.asarray(p1_board)].set(LOC_P0_BOARD + 1)
+    col_top, col_hs, col_len = _columns(p0_board or [], p1_board or [])
+    return state._replace(
+        card_loc=card_loc,
+        col_top=col_top,
+        col_hs=col_hs,
+        col_len=col_len,
+        to_move=jnp.asarray(to_move, dtype=jnp.int8),
+    )
+
+
+def _fill_hand(cards: list[int], reserved: set[int]) -> list[int]:
+    result = list(cards)
+    blocked = set(result) | reserved
+    for card in range(N_CARDS):
+        if len(result) == HAND_SIZE:
+            return sorted(result)
+        if card not in blocked:
+            result.append(card)
+            blocked.add(card)
+    raise AssertionError("could not fill hand")
+
+
+def _columns(p0_board: list[int], p1_board: list[int]):
+    col_top = np.zeros((2, N_COLORS), dtype=np.int8)
+    col_hs = np.zeros((2, N_COLORS), dtype=np.int8)
+    col_len = np.zeros((2, N_COLORS), dtype=np.int8)
+    for player, cards in enumerate([p0_board, p1_board]):
+        for card in cards:
+            color = card // CARDS_PER_COLOR
+            slot = card % CARDS_PER_COLOR
+            col_len[player, color] += 1
+            if slot < 3:
+                col_hs[player, color] += 1
+            else:
+                col_top[player, color] = max(col_top[player, color], slot - 1)
+    return jnp.asarray(col_top), jnp.asarray(col_hs), jnp.asarray(col_len)